Design advanced prompting strategies for production AI agents

Why it matters

A production prompt is a versioned contract that you promote on evidence, and it is never the only control between an attack and the agent.

What to practise

Compare a baseline and a candidate prompt as two Foundry agent versions; the candidate requires a four-field JSON output and treats patient and tool data as untrusted. Complete guardrails at the input, tool call, tool response and output, and confirm the attack cases stop before any conversation or response is created. Promote only if the candidate passes the benign case, keeps both attacks blocked and adds no output failures. Then record a no-go for fine-tuning: the rows have no reviewed target responses.

How it fits

The candidate is 2.1.1’s result contract and 2.1.3’s defensive instructions; the promotion gate is 2.1.4, and the fine-tuning no-go is 2.1.7’s data gate. For 4.2, compare the lab’s pattern-based input and tool-response checks with 4.2.2’s Prompt Shields and 4.2.4’s screening, the output check with 4.2.5’s acceptance gate, and its fail-closed tests with 4.2.7.